[Submitted] Reproducibility of machine learning analyses of 21 cm reionization maps

Machine learning (ML) methods have become popular for parameter inference in cosmology, although their reliance on specific training data can cause difficulties when applied across different data sets. By reproducing and testing networks previously used in the field, and applied to 21cmFast and Simfast21 simulations, we show that convolutional neural networks (CNNs) often learn to identify features of individual simulation boxes rather than the underlying physics, limiting their applicability to real observations. We examine the prediction of the neutral fraction and astrophysical parameters from 21 cm maps and find that networks typically fail to generalise to unseen simulations. We explore a number of case studies to highlight factors that improve or degrade network performance. These results emphasise the responsibility on users to ensure ML models are applied correctly in 21 cm cosmology.